README.md

---
license: cc-by-nc-4.0
task_categories:
- text-generation
language:
- en
size_categories:
- n<1K
---

# SSH Cloze Benchmark

A Cloze-style benchmark for evaluating language models on Social Sciences and Humanities (SSH) text understanding. The benchmark measures whether a model can choose between two equivalent candidate tokens (e.g. *higher* vs. *lower*, *positive* vs. *negative*) in the context of an academic abstract, where the correct choice requires domain knowledge rather than general English fluency.

This dataset was introduced in the technical report *SHARE: Social-Humanities AI for Research and Education* (Gonçalves, de Jager, Knoth, Pride, & Jelicic, 2026) as the evaluation benchmark for the SHARE family of SSH-specialised language models.

## Dataset summary

- **Task:** Cloze-style binary token prediction in academic abstracts.
- **Size:** 275 examples.
- **Fields (disciplines):** 11 SSH fields, 25 examples each — Art, Business, Communication, Economics, Education, Geography, History, Law, Philosophy, Psychology, Sociology.
- **Source:** Out-of-distribution SSH abstracts published in Q1 2026, retrieved from Web of Science and ranked per discipline by citation count. Recency was a requirement in order to minimise risk of training-data contamination for models with earlier cutoffs.
- **Language:** English.
- **Domain:** Social Sciences and Humanities scholarly writing.

## Motivation

Standard LLM benchmarks such as MMLU assume content (often STEM, high-school level) and formats (multiple choice) that are not representative of SSH scholarship, and general perplexity comparisons conflate SSH-specific competence with general English fluency. The SSH Cloze Benchmark isolates SSH-relevant prediction by focusing on tokens where the choice between two equivalent alternatives hinges on domain knowledge. For example, in *"The correlation between social media use and well-being was negative,"* predicting *was* requires only basic English, but predicting *negative* over *positive* requires familiarity with the findings and conventions of SSH literature.

## Data fields

Each row contains:

| Field | Description |
| --- | --- |
| `Record` | Web of Science URL for the source abstract. |
| `Original abstract` | Full unmodified abstract as retrieved from Web of Science. |
| `Cloze abstract` | Abstract rewritten/truncated so that the target token is the final (or otherwise decisive) word, making it suitable for a next-token prediction or masked-token evaluation. |
| `Correct token` | The token the model should prefer, grounded in the original abstract's finding. |
| `Incorrect token` | The equivalent distractor token (same syntactic role, opposite or alternative meaning). |
| `Sign` | `Positive`, `Negative`, or `Neutral` — the direction of the correct token's claim. Distribution: 140 Positive, 97 Negative, 38 Neutral. |
| `Field` | The SSH discipline the abstract belongs to (one of the 11 fields above). |

## Construction

Candidate abstracts were retrieved with a keyword search aimed at finding terms that lend themselves to equivalent-token framing — *positive / negative*, *higher / lower*, *greater / smaller*. Results were ranked by citation count within each discipline, and 25 abstracts were kept per field. Each abstract was then rewritten into a Cloze prompt ending in (or hinging on) the target token, paired with a plausible distractor from the same equivalence class. The most frequent correct tokens are *higher* (31), *lower* (26), *positive* (23), and *negative* (18), followed by a long tail of other comparative and evaluative terms.

## Evaluation

Models are scored on their ability to assign higher probability to the correct token than to the distractor. The report uses **prior-corrected accuracy** to control for the fact that one token in a pair (e.g. *positive* effects) is often more frequent in English than its counterpart, so that models cannot achieve high scores by defaulting to the more common word.

Reported results from the technical report:

| Model | Size | Training tokens | Raw accuracy | Prior-corrected |
| --- | --- | --- | --- | --- |
| Phi-4 | 14B | 9.8T | 81.8% | 81.8% |
| SHARE | 14B | 96B | 77.1% | 79.6% |
| OLMO-2 | 7B | 4T | 78.2% | 76.4% |
| OLMO-2-Step-20k | 13B | 168B | 74.9% | 73.8% |
| Phi-4 | 4B | 5T | 73.8% | 69.8% |
| SHARE | 4B | 28B | 69.8% | 66.2% |
| SSCI-SciBERT-e2 | 110M | ~1B | 66.9% | 67.6% |
| Pythia | 3B | 300B | 65.8% | 63.6% |
| SciBERT | 110M | 3B | 67.9% | 62.9% |
| Pythia | 12B | 300B | 67.3% | 61.5% |
| BERT | 110M | ~5B | 58.2% | 57.5% |

The benchmark is compatible with both causal LMs (scored on next-token logits at the Cloze position) and masked LMs (scored on the masked-token distribution).

## Intended uses

- Comparing SSH-domain competence of causal and masked language models independently of general English proficiency.
- Evaluating domain-specialised pretraining recipes, particularly for social-science and humanities corpora.
- Probing for data contamination: because abstracts are drawn from Q1 2026 publications, pre-2026 models are unlikely to have seen them verbatim.

## Limitations

- **Initial release.** The report describes this as an initial version; the authors plan to expand the number of examples and disciplines.
- **English only.** All abstracts are in English, mirroring the English-centric bias of the SHARE training corpus.
- **Keyword-driven selection.** The requirement that abstracts contain comparative/evaluative keywords (*higher/lower*, *positive/negative*, *greater/smaller*) biases the benchmark toward quantitative or empirically-framed SSH research, and away from purely interpretive humanities writing.
- **Possible LLM contamination in source abstracts.** Since the abstracts are recent, some may themselves have been drafted with LLM assistance.
- **Distractor design.** Distractors are single equivalent tokens; the benchmark does not test open-ended generation, long-range reasoning, or argumentation.
- **Prior correction is necessary.** Because *positive*, *higher*, and *greater* dominate the correct-token distribution, raw accuracy overstates performance; the prior-corrected metric should be the headline number.

## Citation

If you use this dataset, please cite the accompanying technical report:

```
@techreport{goncalves2026share,
  title   = {SHARE: Social-Humanities AI for Research and Education},
  author  = {Gon{\c{c}}alves, Jo{\~a}o and de Jager, Sonia and Knoth, Petr and Pride, David and Jelicic, Nick},
  year    = {2026},
  note    = {arXiv:2604.11152}
}
```

And the original Cloze procedure:

```
Taylor, W. L. (1953). "Cloze procedure": A new tool for measuring readability. Journalism Quarterly, 30(4), 415–433.
```

## License and ethics

Abstracts are drawn from Web of Science-indexed publications. Redistribution should respect publisher terms; the dataset is intended for non-commercial research and evaluation, consistent with the Responsible AI License (RAIL) terms used by the SHARE models.